- Ai refers to many technologies including machine learning and LLMs.
- What follows are thoughts mostly around LLMs and generative AI.
- The biggest innovation around AI is the advance in natural language processing, the ability to query a computer in plain English. This is incredible, and is often overlooked.
- There are domains where LLMs are revolutionary.
- Some of these domains include transcription, data transformation, well-defined programming tasks (such as regex, bash scripts, CSS, and unit tests), and searching a limited corpus.
- All output should be verified.
6a. By human nature, all output will not be verified. - As long as LLMs hallucinate, they should not be relied on for cases where there are stakes and a cost to getting it wrong (e.g. medicine)
- Hallucinations are intrinsic to the technology. LLMs seek patterns and provide likely outcomes. There isn’t a concept of “knowing” or “not knowing”.
8b. There may be a concept of “confidence”, which is strongly inversely correlated to the novelty of the task. - Machine learning is a great tool for detecting patterns that humans cannot see. This is also “AI”, and may have places in (e.g.) medicine.
- No AI should be a first line of defense. See above re: verification. For the example I just gave, machine learning may suggest a cancer. A radiologist should confirm.
- No AI — LLM or ML — has a concept of understanding. There are cases where machine learning has “learned” to look the hospital an X-Ray came from, rather than the X-Ray itself based on patterns in its training data.
11b. This is similar to the way that geo-guesser players will look for artifacts about the Google cars rather than the terrain itself. - AI is not applicable to every domain, including many domains where it is being applied today.
- The fact that AI can create illustrations from words is very cool and novel in a vacuum.
- This fact does not exist in a vacuum, and the technology is built on stolen work.
- The same is true for writing.
15a. It is much harder to notice for most types of writing, and this has fewer defenders.
15b. This is a shame. - Using AI in domains where it is unapt is a crutch and will have consequences, both short- and long-term.
- Among these consequences will be mistakes due to lack of validation, and a defaulting to procedure over understanding. This is a pernicious sort of failure state.
- Like any abstraction, using AI reduces the user’s understanding of what is happening under the hood. This makes debugging hard.
18a. This is not inherently bad. Not everyone needs to understand how or why things work all of the time. Drivers do not need to understand the internal combustion engine. Mechanics should. - Replacing juniors with algorithms disrupts the pipeline that creates experts.
- AI is no replacement for creativity, programmers, writers, or dungeon masters.
20b. I am all of these things, which presents both a vested interest in this belief, as well as qualifications to hold it. - AI can be a useful tool to aid in many tasks. Creativity is the one it is, by nature, least well-suited for. It cannot create what is not in the pattern.
- Using it in these contexts is sad. This may come off as judgmental, which I try not to do. I am comfortable stating that something is lost in shunting creativity, thought, processing, poetry, understanding, art, community, learning to a computer. We are organic, wet, social beings. We crave community, and memories stick best when there is a story attached, an emotion attached. We all remember how mortified we were to confidently mispronounce epitome, hyperbole, or chronic in front of a cute girl. To claim that Cheerios lowers “Chest-er-ol” to a crowd of delighted adults, laughing at your innocent best attempt. This is human, this is brain. Do not give over entirely. How many times have you typed the same word into Google, hoping that this time the definition will stick? This is not learning, but a pale excuse. Even if the machine returns the perfect definition each time, no new wrinkle has been formed in your cerebellum. No new memory has formed. Just mechanical strokes creating output down the line of an infinite telegraph, lost as soon as it was found. There is sadness in this, and I wear my judgment like an albatross. Not proudly, but around by neck for all to see, and all to judge me by. I have sinned by judging, and I may be judged by it.
- No software should be forced upon its users, in any context.
23a. Software that delights will be jumped upon.
23b. Some of these AI tools are the fastest-downloaded apps of all time. This proves some craving. No one is craving integration in every facet.
23c. Security updates and the like can be recommended in the strongest possible terms. If any software must be forced, it is these.
23d. The more seamlessly software is offered, the happier the user is. No one opposes updates. They oppose change and they oppose inconvenience. Minimize both. - AI looks like AL, and we should really do something about serifs. They’re important when 1 l and I can all look the same!
Bonus: AI is not going to destroy civilization, it’s over-invested in, and uses too much energy. But, if it’s the path that gets America on nuclear energy, there’s a silver lining to it all.